Google's Gemini 4 Carbon edges ahead on an AI index, but few people can use it

A one-point lead at the top

On October 28, 2026, Google introduced Gemini 4 Carbon, the successor to Gemini 4 Argon, which became generally available on October 13, 2026. As BornCity reported, Google says Carbon now sits first on the Artificial Analysis Intelligence Index v5.0 with 59 points.

Right behind it are Claude Opus 5.5 with 58 points and GPT-6 Astra with 57, both in their maximum configurations. An index like this blends results from many different tasks into a single number, which makes rankings easy to read but hides how the score was built.

The gap is one point, and no margin of error was published. Without that, it is hard to say whether Carbon is truly ahead or simply tied with its rivals within normal variation.

Built for code and security work

Google designed Carbon for demanding software engineering and cybersecurity jobs. It also aimed the model at long-running tasks. Internally, the company has used it for stretches of days to trace the root causes of bugs in large codebases and fix them.

Google ran internal security tests with the model's protective filters turned off. A benchmark is a standardized test used to compare models. Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench.

On writing secure code, Google says Carbon succeeds more than twice as often as Argon. The Chrome team also reported that Carbon produces high-quality bug fixes more than ten times as often as other leading models. A tool called CodeMender can automate that kind of repair.

All of these figures come from Google or its own teams. Independent confirmation has not appeared yet.

Who can actually use it

Most people cannot try Carbon today. Access begins with selected cybersecurity experts. Trusted partners are meant to follow through a program called Fairwind.

Google gave no date for wider availability or for opening the model to general software developers. It also has not published which organizations qualify. The company is taking part in a voluntary US government program that tests AI models before release.

Restricting a model that is strong at finding and exploiting software flaws is a cautious choice. The same skills that help defenders can also help attackers.

Pricing and the OpenAI release

For developers who reach it through an API (the connection that lets other software call the model), the launch prices match Argon. That is $2 per million input tokens and $10 per million output tokens. A token is a small chunk of text, roughly a word fragment, and is how usage is billed. Cached inputs, meaning text the model has already seen, get a 95% discount.

Google plans to double these to $4 and $20 at a time it has not stated. Anyone budgeting around the launch price should keep that in mind.

Carbon arrived shortly after OpenAI's GPT-6.1 Terra, released October 26, 2026. One later report dated the launch October 27. Terra costs $1 per million input tokens and $5 per million output tokens, and OpenAI says it matches GPT-6 Astra at computer operation for a tenth of the price. On Terminal-Bench 4.0, Terra solves 50.3% of tasks against 21.5% for GPT-5.6 Terra.

How to read the ranking

A narrow lead on a composite index can vanish with the next release from any competitor. The score also averages many kinds of work, and your own work may look nothing like that average.

The more useful comparison is practical. Pick a handful of real tasks you do, run them on the models you can reach, and see which one serves you best. Carbon itself is out of reach for now, so models like those from OpenAI may be the ones you can actually compare.